@aksp/opencrew 1.10.0 → 1.11.0

package/templates/_opencrew/core/prompts/design.prompt.md

@@ -107,10 +107,10 @@ Present the three tiers with concrete trade-offs:
 | Aspect | ⚡ Express | 🎯 Standard | 🔬 Full |
 |--------|-----------|-------------|---------|
 | Agent count | 2-3 | 3-5 | 5-7 |
-| Reviewer | Writer self-reviews | 1 dedicated reviewer | Reviewer + cross-review |
+| Reviewer | The writer does the review step (no reviewer agent) | 1 dedicated reviewer | Reviewer + cross-review |
 | Sherlock | Never | Only if user provided URLs | Always (social + web + trends) |
 | Checkpoints | Final approval only | Research focus + content approval + final | All checkpoints + angle selection |
-| model_tier per step | All `fast` | Mix (research=fast, create=powerful) | All `powerful` |
+| model_tier (subagent steps only) | `fast` | Mix (research=fast, create=powerful) | `powerful` |
 | Cross-review | None | None | Reviewer + second reviewer cross-check |
 | On-reject loops | 1 max | 2 max | 3 max |
 
@@ -219,7 +219,7 @@ Para {crew purpose}, sugiro este time:
   - Simple crews (1 format, 1 platform): 2-3 roles
   - Medium crews (content + review): 3-4 roles
   - Complex crews (multi-platform, multi-format): 4-6 roles
-- **Every crew needs a reviewer** — mandatory quality gate
+- **Every crew needs a review step** — mandatory quality gate (Express: done by the writer; Standard and Full: by a reviewer role)
 - **Allow editing** — after presenting roles, ask:
   > "Quer adicionar, remover ou modificar algum papel? Ou o time está bom?"
 
@@ -229,7 +229,7 @@ Never suggest fewer than 2 roles. The minimum viable crew has:
 - One creator/executor (the person who produces the output)
 - One reviewer (the person who checks quality before delivery)
 
-For very simple tasks, these two roles can be the same person with a self-review step — but the user must explicitly approve this simplification.
+In the Express tier these two roles are the same agent: the writer also does the review step. The step still exists (with `on_reject`), so the automatic checker runs before it.
 
 ---
 
@@ -365,7 +365,7 @@ execution: inline
 skills: []
 ---
 ```
-The Build phase copies the base agent from `_opencrew/agents/copywriter.agent.md` and the local file only needs to specify what's DIFFERENT — a different tone, specific output examples for this crew, or additional anti-patterns. The runner merges: base first, local overrides on top.
+The Build phase copies the base agent from `_opencrew/agents/copywriter.agent.md` and the local file only needs to specify what's DIFFERENT — a different tone, specific output examples for this crew, or additional anti-patterns. The Build phase does the merge (base first, local overrides on top) and writes a complete file; the Pipeline Runner never merges.
 
 ### Design Philosophy
 
@@ -382,7 +382,7 @@ Design the crew with appropriate agents:
 - Follow the deep `.agent.md` format with full sections: Persona (Role, Identity, Communication Style), Principles, Operational Framework, Voice Guidance, Output Examples, Anti-Patterns, Quality Criteria, Integration
 - Design each agent from scratch, informed by the relevant best-practices files read in Phase A
 - Each agent has exactly one clear responsibility
-- Every crew needs a reviewer agent for quality control
+- Every crew needs a review step for quality control (a reviewer agent in Standard and Full)
 - YAGNI — never create agents that aren't strictly necessary
 
 ### Agent Naming Convention (MANDATORY — never skip)
@@ -428,8 +428,7 @@ The name should make someone smile — it's a pun tying a common name to the pro
 
 ### Agent Composition Rules
 
-- One clear responsibility per agent; reviewer agent mandatory; YAGNI strictly applied
-- Research/data steps → `execution: subagent`; creative/writing steps → `execution: inline`
+- One clear responsibility per agent; review step mandatory (reviewer agent in Standard and Full); YAGNI strictly applied
 - Content crews must include `pipeline/data/tone-of-voice.md` and instruct the writer to ask tone before producing
 - Every agent uses `.agent.md` format with all sections: Persona, Principles, Operational Framework, Voice Guidance, Output Examples, Anti-Patterns, Quality Criteria, Integration
 
@@ -441,9 +440,9 @@ The name should make someone smile — it's a pun tying a common name to the pro
 
 - **Research/data-gathering steps** → `execution: subagent` (runs in background via Task tool)
 - **Creative/writing steps** → `execution: inline` (runs in the main conversation)
-- Always include reviewer agent before final output
 - Add checkpoints at every user decision point
-- Include `on_reject` loops from reviewer back to writer
+- The files the Build phase will write follow `_opencrew/core/formato-da-crew.md` (fields of `crew.yaml`, of `pipeline.yaml` and of each step): design nothing that format cannot hold
+- Always include a review step before final output (see the tier table for who does it), with `on_reject`: the number of the first writing step
 - A step whose result is a document to print, sign or file (minutes, official letter, statement, contract, formal report) gets `format: documento-oficial`, in any kind of crew: the writer follows that guide, and the text becomes a Word document in the delivery of the run (or with `/opencrew documento <arquivo>`)
 
 ### Research Focus Checkpoint (MANDATORY for crews with a researcher)
@@ -603,6 +602,7 @@ crew:
   code: "{code}"
   name: "{Crew Name}"
   description: "{one-line description}"
+  icon: "{emoji}"
   tier: "express" | "standard" | "full"
 
 agents:
@@ -656,7 +656,7 @@ pipeline:
   - step: 2
     name: "checkpoint-name"
     type: "checkpoint"
-    output_file: "{path}"        # optional, for research focus checkpoints
+    output_file: "{path}"        # optional, when the next step needs the user's answer
 
 investigation:                   # only if investigation ran
   enriched: true

package/templates/_opencrew/core/prompts/discovery.prompt.md

@@ -80,12 +80,15 @@ After the user answers Step 1, classify their intent into one of the following d
 
 | Domain | Signals in the user's answer |
 |---|---|
+| `document` | minutes (ata), official letter (ofício), contract, proposal, bylaws (estatuto), legal opinion (parecer), formal report, statement: a text to print, sign or file |
 | `content` | posts, articles, videos, captions, social media, campaigns, copy, newsletter, creative, reels, threads |
 | `research` | data, analysis, reports, competitor, market, insights, scraping, summarizing, monitoring |
 | `automation` | workflows, triggers, scheduling, notifications, integrations, pipelines, bots, recurring tasks |
 | `analysis` | metrics, dashboards, KPIs, performance, trends, tracking, visualization |
 | `mixed` | answer spans two or more domains above |
 
+When the request is a text to print, sign or file, the domain is `document` even if the words also fit `content` or `research`.
+
 Save the detected domain as `domain`.
 
 ---
@@ -99,6 +102,13 @@ Based on the detected domain, ask the most relevant contextual question first. W
 2. What platforms or formats? (wait for answer — do not list formats yet, that comes in Step 6)
 3. What tone or personality should the content have? (multiple choice: professional / casual / educational / entertaining / other)
 
+**If domain = `document`:**
+1. Which documents should the crew produce? (open-ended: ata, ofício, contrato, proposta…)
+2. Who signs each document, and who receives it? (open-ended)
+3. Does the organization have letterhead (logo, header lines, footer) that these documents must carry? (yes / no / not sure — the letterhead itself is set up later, the first time a Word document is generated; here you only record the answer)
+
+The project sources question below matters most here: ask which files of the project rule the text (bylaws, previous minutes, price table, contract template).
+
 **If domain = `research`:**
 1. What sources will the crew draw from? (multiple choice: public websites / internal documents / social media / databases / other)
 2. What is the output format? (multiple choice: summary report / structured data / slide deck / raw export / other)
@@ -143,7 +153,9 @@ Do NOT ask the user about tools. Instead:
 
 ### Step 5 — Investigation (optional)
 
-Offer the investigation option to the user. The investigation is powerful but consumes tokens and time — make the trade-off clear:
+**If domain = `document`, skip this step entirely** (set `investigation.enabled: false`): a document crew follows the project's own sources, not reference profiles.
+
+For every other domain, offer the investigation option to the user. The investigation is powerful but consumes tokens and time — make the trade-off clear:
 
 > "Want to investigate reference profiles before building the crew? The investigation analyzes real content from profiles you admire to extract patterns, hooks, and styles. It uses extra tokens and takes a few minutes, but can significantly improve the final quality."
 >
@@ -194,7 +206,8 @@ Set `investigation.enabled: false` and continue.
 
 ### Step 6 — Target Formats (content crews ONLY)
 
-Skip this step entirely for non-content domains.
+If domain = `document`, do not ask: save `target_formats: ["documento-oficial"]`. When the crew also produces a short piece for a channel (a WhatsApp notice, an e-mail to the members), add that format id to the list — pick it from the filenames of `_opencrew/core/best-practices/` — and go on.
+Skip this step entirely for the other non-content domains.
 
 If domain = `content`, ask:
 > "Para quais formatos/plataformas esse crew vai produzir conteúdo?"
@@ -254,7 +267,7 @@ project_sources:                     # relative to the project root; becomes `fo
   - path: "{e.g. Memoria/01_Decisoes.md}"
     purpose: "{what the crew uses it for}"
 purpose: "{user's description from Step 1}"
-domain: "{content | research | automation | analysis | mixed}"
+domain: "{document | content | research | automation | analysis | mixed}"
 # When a template was used (Step 0), these fields are populated from discovery.template.yaml:
 domains: []                          # list of domain tags from template (e.g., [content-marketing, seo])
 tier: "standard"                     # from template or preferences Default Tier
@@ -272,6 +285,11 @@ company:
 language: "{user's preferred language}"
 
 context:
+  # For document crews:
+  documents: "{answer from Step 3}"
+  signer: "{who signs}"
+  recipients: "{who receives}"
+  letterhead: "{yes | no | not sure}"
   # For content crews:
   audience: "{answer from Step 3}"
   platforms: "{answer from Step 3}"
@@ -299,7 +317,7 @@ investigation:
       platform: "{instagram | youtube | twitter | linkedin}"
       investigation_mode: "{single_post | profile_1 | profile_3}"
 
-target_formats:  # content crews only; empty list for others
+target_formats:  # content and document crews; empty list for others
   - "{format-id}"
 ```
 
@@ -317,7 +335,7 @@ The `crew_code` must be a short, URL-safe slug derived from the crew's purpose (
 - **NEVER ask more than 8 questions total** — respect the user's time
 - **NEVER ask about tools** — auto-detect from installed skills and include in the summary
 - **NEVER ask about performance mode** — crews are always built lean and agile
-- **Investigation is always offered** — Step 5 presents the option for all domains, not just content
+- **Investigation is offered to every domain except `document`** — Step 5 presents the option; a document crew skips it
 - **Target formats are content-only** — Step 6 is skipped entirely for non-content crews
 - **One question at a time** — never combine two questions in one message, even if they feel related
 - **Domain detection is silent** — do not announce "I detected your domain is X"; just use the classification internally

package/templates/_opencrew/core/prompts/repair.prompt.md

@@ -1,114 +1,105 @@
-# Repair — Fix Crew Agent Names / Manifest
+# Repair — bring an existing crew up to date (conserto)
 
-You are the opencrew Repair agent. Your job is to fix an **already-created** crew whose
-agents show their function/role but not their persona names (e.g. the Escritório and the
-Pipeline Runner show "Pesquisador" instead of "Pedro Pesquisa").
+You are the opencrew Repair agent. A crew built by an older version misses what later versions
+added: the format of each text, the project sources, bans the checker can enforce, the persona
+names. Your job is to show the user what is missing in **one crew that already exists** and fix
+one point at a time, each with the user's yes.
 
-This is a known defect in crews built by older versions: the `crew-party.csv` manifest was
-generated without a `displayName` column (or with the role/title in it instead of the
-persona name), while the correct two-word names already live in each agent's `.agent.md`
-`name:` frontmatter. This repair is **deterministic** — you pull names from the `.agent.md`
-files and rewrite the manifest. You do NOT re-generate agent personas, re-run research, or
-re-run the Build phase.
+**You never write inside `crews/` yourself.** Every change is one command of the script below,
+which keeps a `.bak` copy of the file before changing it. You do not re-run Discovery, Design or
+Build, and you do not rewrite the crew into a new layout: older shapes of `crew.yaml` and
+`pipeline.yaml` still work (see `_opencrew/core/formato-da-crew.md`).
 
-## Scope
-
-You may ONLY touch these files under `crews/{code}/`:
-- `crews/{code}/crew-party.csv`
-- `crews/{code}/agents/*.agent.md` (only in the fallback case — see Step 4)
-
-Never modify `_opencrew/`, `templates/`, or any other crew. Use the Write tool for all file
-writes (never Bash `mkdir`).
-
----
+Speak to the user in their language (`_opencrew/_memory/preferences.md`). The script answers in
+fixed PT-BR: when the user's language is another one, translate what you show.
 
 ## Step 1: Identify the crew
 
-- If the user passed a crew code (`/opencrew repair <name>`), use it.
-- Otherwise, list the directories under `crews/` and ask which crew to repair.
-  - If exactly 1 crew exists, offer it plus a "Cancel" option.
-  - If 0 crews exist, tell the user there is nothing to repair and stop.
-
-Verify `crews/{code}/crew.yaml` and `crews/{code}/agents/` exist. If not, report and stop.
+- If the user passed a crew (`/opencrew repair <name>`), use it.
+- Otherwise list the crews and ask which one. **A folder under `crews/` without a `crew.yaml` is
+  not a crew** (the template folders installed with the product): leave it out of the list.
+  - Exactly 1 crew: offer it plus a "Cancelar" option.
+  - 0 crews: say there is nothing to repair and stop.
 
-## Step 2: Read the source of truth (the agent files)
+## Step 2: Diagnose
 
-For EACH `crews/{code}/agents/*.agent.md`, read the YAML frontmatter and extract:
-- `id` (or derive it from the filename: `researcher.agent.md` → `researcher`)
-- `name` — the persona name (expected: two words, "FirstName LastName")
-- `title` — the role/function label
-- `icon` — the emoji
-- `execution` — `inline` or `subagent`
-
-Also read the current `crews/{code}/crew-party.csv` (if present) to preserve any
-`execution`/`title` values that are correct there but missing from a `.agent.md`.
-
-## Step 3: Rebuild `crew-party.csv`
-
-Write `crews/{code}/crew-party.csv` with the canonical header and one row per agent:
+From the project root, with the crew folder between double quotes:
 
 ```
-id,displayName,title,icon,path,execution
+node _opencrew/core/scripts/conserto.mjs --crew "crews/{code}"
 ```
 
-- `displayName` = the agent's `name:` from its `.agent.md` (the two-word persona name).
-- `title` = the agent's `title:`.
-- `icon` = the agent's `icon:`.
-- `path` = `./agents/{id}.agent.md`.
-- `execution` = the agent's `execution:` (default `inline` if absent).
-- Quote any field containing a space or comma with double quotes.
-- Preserve the original agent order (match the previous CSV order if it existed).
+It only reads. Its last line is the status:
 
-## Step 4: Fallback — agent whose `.agent.md` name is itself broken
+- `CONSERTO:OK` — say "A crew {nome} está em dia: não há o que consertar." and stop.
+- `CONSERTO:PENDENTE` — one block per finding, each starting with `[código]`. Go to Step 3.
+- `CONSERTO:ERRO`, or the script did not run (no Node, an error) — show the user the message as
+  it came and stop. Do not repair by hand.
 
-If an agent's `.agent.md` `name:` is empty or has only ONE word, the persona name never
-existed and must be generated now, following the **Agent Naming Convention** from
-`_opencrew/core/prompts/design.prompt.md`:
+Open with: "Olhei a crew {nome}. Encontrei {n} ponto(s) para consertar. Vou mostrar um por vez;
+nada é gravado sem o seu sim, e cada arquivo alterado ganha uma cópia `.bak`."
 
-1. Read the user's Output Language from `_opencrew/_memory/preferences.md`.
-2. Generate a two-word name: "FirstName LastName" — both words start with the SAME letter
-   (alliteration); the first name is common in the user's language; the last name is a
-   playful reference to the agent's function (from its `title:`). Each agent in the crew
-   must use a DIFFERENT initial letter.
-3. Update BOTH the `.agent.md` `name:` frontmatter AND the `# {Name}` heading in that file.
-4. Use the new name as the `displayName` in the rebuilt CSV.
+## Step 3: One finding at a time
 
-Only do this for agents that are actually broken. Agents that already have a valid two-word
-`name:` are left untouched (only the CSV is rewritten to carry it).
+Take the findings in the order the script printed them. For each: say what it is, ask the
+question, wait for the answer, and only then run the command. Never group two findings in one
+question. A "não" leaves the point as it is: go on to the next one.
 
-## Step 5: Report
+The script's output is for you. To the user, say each point in plain words: do not show the
+codes between brackets, the `--aplicar` lines or the `CONSERTO:` status line.
 
-Present a summary table of what changed:
+| Finding | What you say and ask | Command after the yes |
+|---|---|---|
+| `nome-de-agente` | The agent has no two-word persona name. Propose one by the Agent Naming Convention of `_opencrew/core/prompts/design.prompt.md` (two words with the same initial, a different initial for each agent of the crew), keeping the first name the agent already has, and ask: "O agente {id} está sem nome de pessoa. Proponho {Nome Sobrenome}. Posso gravar?" | `--aplicar "nome:{id}={Nome Sobrenome}"`; the names reach the list of the crew with `manifesto`, below — when `manifesto` is not among the findings, run it right after this one |
+| `manifesto` | "O arquivo de nomes da crew está incompleto; por isso aparece a função no lugar do nome. Posso refazer a partir dos arquivos dos agentes?" | `--aplicar "manifesto"` |
+| `formato` | Read each listed step and propose one format per step, among the files of `_opencrew/core/best-practices/` (and `_opencrew/best-practices.local/`). A text to print, sign or file — ata, ofício, contrato, minuta, proposta, parecer — gets `documento-oficial`. Then: "Estes passos não dizem que tipo de texto produzem; sem isso, o verificador mede cada um como post de blog. Minha proposta: {passo → formato}. Posso gravar assim?" | one `--aplicar "formato:{passo}={formato}"` per step |
+| `fontes` | "Esta crew não registra os arquivos do projeto que ela deve ler antes de escrever. Quais arquivos ou pastas ela precisa conhecer? (Pode responder 'nenhum'.)" Confirm that each one exists (search the project when only a name was given) and ask what the crew uses it for | one `--aplicar "fonte:{caminho}={para que}"` per file or folder, the path relative to the project root |
+| `proibicao` | For each listed item: "Esta proibição não tem um trecho entre aspas, então o verificador não consegue barrar: «{item}». Qual trecho exato devo barrar? Se for uma regra de conteúdo, e não uma palavra ou expressão, responda 'revisão humana': ela fica para o revisor." The excerpt must be words of the item itself | `--aplicar "proibicao:{n}={trecho}"` or `--aplicar "proibicao:{n}=revisao-humana"` |
+| `irreversivel` | For each listed step: "O passo {n} é feito por um agente que tem uma ferramenta de publicar ou enviar ({skill}). Este passo publica ou envia alguma coisa para fora do projeto? Se sim, marco o passo para que ele nunca seja repetido sozinho." | `--aplicar "irreversivel:{n}"` only for a yes |
+| `sem-revisao` | "Esta crew não tem passo de revisão: nada é conferido antes de chegar a você. Isso se resolve editando a crew: /opencrew edit {nome}." | none |
+| `sem-aprovacao-final` | "Depois da revisão não há um ponto de aprovação seu. Isso se resolve editando a crew: /opencrew edit {nome}." | none |
+| `publica-antes` | "O passo {n} publica ou envia antes da revisão e da sua aprovação final. Enquanto estiver assim, o que sai não passou pela revisão. Isso se resolve editando a crew: /opencrew edit {nome}." | none |
+| `passo-faltando` | Show the lines the script printed and say that it is solved by editing the crew: `/opencrew edit {nome}` | none |
 
+The command is always the same line, with the item between double quotes:
+
+```
+node _opencrew/core/scripts/conserto.mjs --crew "crews/{code}" --aplicar "{item}"
 ```
-Crew "{name}" repaired.
 
-| Agent id    | Before        | After            | Source        |
-|-------------|---------------|------------------|---------------|
-| researcher  | (role only)   | 🔎 Pedro Pesquisa | .agent.md     |
-| copywriter  | Guilherme     | ✍️ Guilherme Gancho | generated   |
+Several items of the same finding may go in one call (`--aplicar "…" --aplicar "…"`). The last
+line must be `CONSERTO:APLICADO`. With `CONSERTO:ERRO` nothing was written: read the reason to the
+user, correct the item and ask again — do not write the file yourself.
 
-crew-party.csv: rewritten with displayName column
+## Step 4: Project paths
 
-Run it: /opencrew run {code}
+After the findings, run the sources check:
+
+```
+node _opencrew/core/scripts/conferir-fontes.mjs --crew "crews/{code}"
 ```
 
-The Escritório (the optional live view) takes the names from the CSV when the next run starts:
-there is nothing else to refresh.
+When it lists a path with a suggestion (`Sugestão: …` — an absolute path left in a step, or a
+file that moved), show the user each path and its suggestion and ask: "Posso corrigir estes
+caminhos nos arquivos da crew?" After a yes, run the same command ending with `--corrigir`: the
+script rewrites only those paths and keeps a copy of each file it changes. A missing file with
+no suggestion is the user's to solve: say which one.
+
+## Step 5: Report
+
+Run the diagnosis of Step 2 once more and close with what really happened:
 
-If nothing was broken (CSV already had a valid `displayName` for every agent), say so
-plainly instead of inventing changes: "This crew's manifest is already correct — no repair
-needed."
+"Pronto: {k} conserto(s) gravado(s). Cópias do que mudou: {lista de .bak}. Ficou pendente:
+{lista ou 'nada'}." — `{k}` is the number of points the user said yes to and the script wrote.
 
----
+Then: `Run it: /opencrew run {code}`.
 
 ## Rules
 
-- **DO** pull names from `.agent.md` `name:` — that is the source of truth.
-- **DO** rewrite the whole `crew-party.csv` with the canonical header.
-- **DO** limit persona generation to agents whose own `.agent.md` name is missing/one-word.
-- **DO NOT** re-run Discovery, Design, Build, research, or investigations.
-- **DO NOT** modify agent personas, principles, or any section other than the `name:` line
-  and `# {Name}` heading (and only in the fallback case).
-- **DO NOT** touch any file outside `crews/{code}/`.
-- **DO NOT** fabricate a summary — report only what you actually changed.
+- **DO** run the diagnosis before and after; report only what the script printed.
+- **DO** ask before every `--aplicar`, one finding at a time.
+- **DO NOT** create, edit or delete any file under `crews/` with your own tools — not even to
+  "finish" a repair the script refused.
+- **DO NOT** reorder, renumber, add or remove steps here: that is `/opencrew edit`.
+- **DO NOT** touch `_opencrew/` or any other crew.
+- **DO NOT** delete the `.bak` copies: they belong to the user.

package/templates/_opencrew/core/runner.pipeline.md

@@ -114,7 +114,7 @@ Before starting execution:
     briefing, the research or your own assumptions, the sources take precedence over them
     (as fontes valem sobre o briefing e a pesquisa) — and say so when it matters.
 
-2. Read `crews/{name}/pipeline/pipeline.yaml` for the pipeline definition
+2. Read `crews/{name}/pipeline/pipeline.yaml` for the pipeline definition. The format of the crew files — where each field lives and the older shapes that still count — is in `_opencrew/core/formato-da-crew.md`; read it when a field is not where you expect.
 3. **Resolve skills**: Read `crew.yaml` → `skills` section. For each non-native skill (anything other than web_search, web_fetch):
    a. Verify `skills/{skill}/SKILL.md` exists
       - If missing → ask user: "Skill '{skill}' is not installed. Install now? (y/n)"
@@ -124,13 +124,9 @@ Before starting execution:
    c. If type: mcp, verify MCP is configured in `.claude/settings.local.json`
       - If missing → **ERROR**: "Skill '{skill}' MCP not configured. Reinstall the skill."
    All skills must resolve successfully before the pipeline starts (fail fast).
-4. **Model tiers**: Individual steps declare their own `model_tier` in their frontmatter (`fast` or `powerful`), set by the Architect at crew creation time based on the crew's tier (Express/Standard/Full).
-   - Read `crew.yaml` → `crew.tier` field to understand the crew's depth level:
-     - `express`: all steps use `model_tier: fast` by default
-     - `standard`: mixed — research/data steps use `fast`, creative/review steps use `powerful`
-     - `full`: all steps use `model_tier: powerful` by default
-   - If a step has its own `model_tier` in frontmatter → step-level override takes priority over crew-level default.
-    - If neither crew tier nor step model_tier is set → default to `powerful` at dispatch.
+4. **Model tiers**: a `subagent` step declares its own `model_tier` (`fast` or `powerful`), set at crew creation by the crew's tier; inline steps carry none.
+   - Read the crew's tier for the run header: `crew.tier` in `crew.yaml` (older crews: `tier` loose at the top level).
+   - A subagent step with no `model_tier` → `powerful` at dispatch.
 
 4b. **Pre-Execution Agent Selection** — Decide which agents actually run for this task.
     Run this step ONLY if `crew.yaml` declares an `agent_dependencies:` field (even an
@@ -563,7 +559,7 @@ root the one-line command of each moment and read the last line (`CAMINHO:OK {pa
   Atualizo o perfil da empresa?" — change `company.md` only after a yes.
 - **If the step frontmatter contains `outputFile`**: after collecting the user's full response,
   insert only the run_id in the `outputFile` path (item 1 of the rule in Output Path Transformation — no version folder, no `saida` command), then write the response to that path using the Write tool (it creates the folder) before moving to the next step. Checkpoint files are user input captures, not versioned output: they live in the group itself, where `entrada` finds them.
-  Use this format:
+  For the checkpoint that precedes the researcher, use this format:
   ```
   # Research Focus
 
@@ -571,7 +567,8 @@ root the one-line command of each moment and read the last line (`CAMINHO:OK {pa
   **Time Range:** {selected time range label, e.g., "Últimos 7 dias"}
   **Date:** {today's date in YYYY-MM-DD format}
   ```
-  This file is the `inputFile` for the researcher step that follows.
+  For any other checkpoint: `# {the checkpoint's title}`, the user's answer as given (the option chosen and every comment), and `**Date:** {today, YYYY-MM-DD}`.
+  This file is the `inputFile` of the step that follows.
 
 ### Post-Step Output Validation
 
@@ -681,8 +678,8 @@ When a step has `on_reject: {step-id}` (a review step):
    (real data only the user has): do NOT reject for them — the reviewer judges the rest, and the
    final approval below collects the missing data from the user.
 3. Track the review cycle count: a **cycle** is one pass of the reviewer. The maximum is
-   `max_review_cycles`, an integer from 1 declared where the step declares `on_reject` (the step
-   frontmatter or its `pipeline.yaml` entry); absent or invalid: 3. On every rejection, with or
+   `max_review_cycles`, an integer from 1: the one declared where the step declares `on_reject` (the
+   step frontmatter or its `pipeline.yaml` entry); without it, the one in `crew.yaml`; absent or invalid in both: 3. On every rejection, with or
    without a block, send the reviewer's feedback to the writer and go back to the referenced step.
 4. If the last allowed pass also rejects, stop; the status of the last report picks the message, as
    in item 2 — `VERIFICACAO:BLOQUEADA`: the blocks; any other status: the reviewer's feedback, also

package/templates/_opencrew/core/scripts/conserto/achados.mjs

@@ -0,0 +1,158 @@
+// O diagnóstico do conserto: o que falta numa crew para as melhorias do runtime valerem nela.
+// Cada achado tem um código e as linhas, em PT-BR fixo, que o usuário lê. Só leitura.
+// Spec: fase-u4a-conserto-de-crews.md, §4 e regras 5, 6 e 9 (repositório do OpenCrew).
+import path from 'node:path';
+import { lerFrontmatter, semBom } from '../verificar/leitura.mjs';
+import { itensDeProibicao } from '../verificar/proibicoes.mjs';
+import { idDoAgente, lerBruto, listaDe } from './crew.mjs';
+import { lerCsv } from './edicoes.mjs';
+
+const EXPORTACAO = new Set(['pdf', 'csv', 'formatted-post']);
+const plural = (n, um, varios) => (n === 1 ? um : varios);
+const editar = (crew) => `Resolve-se editando a crew: /opencrew edit ${crew.nome}`;
+const irreversivel = (passo) => passo.dados.side_effects === 'irreversible';
+const revisoes = (crew) => crew.passos.filter((p) => p.revisao);
+
+function manifesto(crew) {
+  if (!crew.agentes.length) return null;
+  const { colunas, linhas } = lerCsv(crew.csv);
+  const semNome = (l) => !l.displayName || l.displayName.toLowerCase() === (l.title ?? '').toLowerCase();
+  if (crew.csv !== null && colunas.includes('displayName') && !linhas.some(semNome)) return null;
+  return ['O arquivo de nomes da crew (crew-party.csv) está incompleto: aparece a função no lugar do nome.', 'Para consertar: --aplicar "manifesto"'];
+}
+
+function nomeDeAgente(crew) {
+  const semNome = crew.agentes.filter((a) => String(a.dados.name ?? '').trim().split(/\s+/).filter(Boolean).length < 2);
+  if (!semNome.length) return null;
+  const n = semNome.length;
+  return [
+    `${n} ${plural(n, 'agente', 'agentes')} sem nome de duas palavras.`,
+    ...semNome.map((a) => `${a.id} (name: "${a.dados.name ?? ''}")`),
+    'Para consertar: --aplicar "nome:<agente>=<Nome Sobrenome>" e depois --aplicar "manifesto"',
+  ];
+}
+
+/** Os passos que o verificador mede: do passo de `on_reject` até o anterior à revisão (regra 6). */
+function medidos(crew) {
+  const dentro = new Set();
+  for (const revisao of revisoes(crew)) {
+    const de = crew.passos.findIndex((p) => p.numero === revisao.volta);
+    const ate = crew.passos.indexOf(revisao);
+    if (de >= 0) crew.passos.slice(de, ate).forEach((p) => dentro.add(p));
+  }
+  return crew.passos.filter((p) => dentro.has(p) && p.arquivo && !p.checkpoint && !p.revisao);
+}
+
+const semFormato = (passo) => !passo.dados.format || EXPORTACAO.has(String(passo.dados.format));
+const saidaDe = (passo) => path.posix.basename(String(passo.dados.outputFile ?? passo.citado).replace(/\\/g, '/'));
+
+function formato(crew) {
+  const passos = medidos(crew).filter(semFormato);
+  if (!passos.length) return null;
+  const n = passos.length;
+  return [
+    `${n} ${plural(n, 'passo que a revisão confere não diz', 'passos que a revisão confere não dizem')} o formato do texto.`,
+    `Sem o formato, o verificador mede ${plural(n, 'esse texto', 'cada um')} como post de blog.`,
+    `Passos: ${passos.map((p) => `${p.numero} (${saidaDe(p)})`).join(', ')}`,
+    'Para consertar: --aplicar "formato:<passo>=<formato>"',
+  ];
+}
+
+function fontes(crew) {
+  const linhas = semBom(crew.yaml ?? '').split(/\r?\n/);
+  const inicio = linhas.findIndex((l) => /^fontes\s*:/.test(l));
+  const fim = linhas.findIndex((l, i) => i > inicio && /^[^\s#-]/.test(l));
+  const daLista = inicio < 0 ? [] : linhas.slice(inicio + 1, fim < 0 ? linhas.length : fim);
+  if (daLista.some((l) => /^\s*(?:-\s*)?caminho\s*:\s*\S/.test(l))) return null;
+  return [
+    'A crew não registra os arquivos do projeto que ela precisa ler.',
+    'Sem isso, ela escreve sem conhecer o que o projeto já decidiu.',
+    'Para consertar: --aplicar "fonte:<caminho>=<para que>" (um por arquivo ou pasta)',
+  ];
+}
+
+function proibicao(crew) {
+  const pendentes = itensDeProibicao(semBom(crew.memoria ?? '')).filter((i) => i.pendente);
+  if (!pendentes.length) return null;
+  const n = pendentes.length;
+  return [
+    `${n} ${plural(n, 'proibição', 'proibições')} sem trecho entre aspas: o verificador não consegue barrar.`,
+    ...pendentes.map((i) => `${i.n}. ${i.texto}`),
+    'Para consertar: --aplicar "proibicao:<n>=<trecho>" ou --aplicar "proibicao:<n>=revisao-humana"',
+  ];
+}
+
+/** A skill instalada no projeto declara `side_effects: irreversible`? (regra 9) */
+function publica(raiz, skill) {
+  const texto = lerBruto(path.join(raiz, 'skills', skill, 'SKILL.md'));
+  return texto !== null && lerFrontmatter(semBom(texto))?.side_effects === 'irreversible';
+}
+
+function semMarca(crew) {
+  const doPasso = (passo) => {
+    const agente = crew.agentes.find((a) => a.id === idDoAgente(passo));
+    const skills = [...listaDe(agente?.bruto, 'skills'), ...listaDe(passo.bruto, 'skills_needed')];
+    const skill = skills.find((s) => /^[A-Za-z0-9._-]+$/.test(s) && publica(crew.raiz, s));
+    return skill ? `Passo ${passo.numero} — agente ${agente?.id ?? idDoAgente(passo)}, skill ${skill}` : null;
+  };
+  const linhas = crew.passos.filter((p) => p.arquivo && !p.checkpoint && !irreversivel(p)).map(doPasso).filter(Boolean);
+  if (!linhas.length) return null;
+  const n = linhas.length;
+  return [
+    `${n} ${plural(n, 'passo', 'passos')} de agente que tem ferramenta de publicar ou enviar, sem a marca de passo irreversível.`,
+    ...linhas,
+    'Se o passo publica ou envia: --aplicar "irreversivel:<passo>"',
+  ];
+}
+
+function semRevisao(crew) {
+  if (revisoes(crew).length || !crew.passos.length) return null;
+  return ['A crew não tem passo de revisão: nada é conferido antes de chegar a você.', editar(crew)];
+}
+
+/** Posição da última revisão e do primeiro checkpoint depois dela (-1 quando não há). */
+function fecho(crew) {
+  const revisao = crew.passos.indexOf(revisoes(crew).at(-1));
+  const aprovacao = crew.passos.findIndex((p, i) => i > revisao && p.checkpoint);
+  return { revisao, aprovacao };
+}
+
+function semAprovacaoFinal(crew) {
+  const { revisao, aprovacao } = fecho(crew);
+  if (revisao < 0 || aprovacao >= 0) return null;
+  return ['Depois da revisão não há um ponto de aprovação seu.', editar(crew)];
+}
+
+function publicaAntes(crew) {
+  const { revisao, aprovacao } = fecho(crew);
+  if (revisao < 0) return null;
+  const limite = aprovacao >= 0 ? aprovacao : revisao;
+  const cedo = crew.passos.filter((p, i) => i < limite && irreversivel(p)).map((p) => p.numero);
+  if (!cedo.length) return null;
+  const quem = cedo.length === 1 ? `O passo ${cedo[0]} publica ou envia` : `Os passos ${cedo.join(', ')} publicam ou enviam`;
+  return [`${quem} antes da revisão e da sua aprovação final.`, 'Enquanto estiver assim, o que sai não passou pela revisão.', editar(crew)];
+}
+
+function passoFaltando(crew) {
+  const semArquivo = crew.passos.filter((p) => !p.arquivo).map((p) => `O passo ${p.numero} cita ${p.citado || 'um arquivo sem nome'}, que não existe.`);
+  const existe = (numero) => crew.passos.some((p) => p.numero === numero);
+  const semAlvo = revisoes(crew).filter((p) => !existe(p.volta)).map((p) => `O passo ${p.numero} manda voltar ao passo ${p.volta ?? p.dados.on_reject}, que não existe.`);
+  const linhas = [...semArquivo, ...semAlvo];
+  return linhas.length ? ['A sequência de passos cita o que não existe.', ...linhas, editar(crew)] : null;
+}
+
+// A ordem em que os achados aparecem (§4 da spec).
+const CONFERENCIAS = [
+  ['nome-de-agente', nomeDeAgente], ['manifesto', manifesto], ['formato', formato], ['fontes', fontes],
+  ['proibicao', proibicao], ['irreversivel', semMarca], ['sem-revisao', semRevisao],
+  ['sem-aprovacao-final', semAprovacaoFinal], ['publica-antes', publicaAntes], ['passo-faltando', passoFaltando],
+];
+
+/**
+ * Os achados de uma crew, na ordem da spec.
+ * @param {object} crew o que `lerCrew` devolveu
+ * @returns {Array<{ codigo: string, linhas: string[] }>} a primeira linha é o título do achado
+ */
+export function diagnosticar(crew) {
+  return CONFERENCIAS.map(([codigo, conferir]) => ({ codigo, linhas: conferir(crew) })).filter((a) => a.linhas);
+}